PETROLEUM EXPLORATION AND DEVELOPMENT

Volume 39, Issue 3, June 2012

Online English edition of the Chinese language journal

Cite this article as: PETROL. EXPLOR. DEVELOP., 2012, 39(3): 349361. RESEARCH PAPER

Schemes for automatic history matching of reservoir

modeling: A case of Nelson oilfield in UK
KAZEMI Alireza*, STEPHEN Karl D
Institute of Petroleum Engineering, Heriot-Watt University, Edinburgh EH14 4AS, UK

Abstract: Schemes for automatic history matching of reservoir modeling are studied for the Nelson oilfield in the Central North Sea. A
complete workflow of automatic history matching involves selection of reservoir variables that require modification and parameter up-
dating schemes, automatic history matching, data analysis, and combination of the best results to obtain an ensemble of best reservoir
models. Automatic history matching of Nelson field is conducted using production and time-lapse seismic data, with global sin-
gle-variable approach, regional multi-variable approach and local multi-variable approach as updating schemes, net to gross, horizontal and
vertical permeability as updating parameters. It is revealed that local multi-variable approach can effectively improve history matched re-
sults by reducing the number of simulation models, saving computing time and increasing the simulation precision. Global single-variable
approach is only a suitable parameter updating scheme for cases where the history matching parameters are independent. Regional
multi-variable approach is suitable for the cases where there is strong dependency between properties chosen for updating, and there are
wells very close together with strong interaction. Local multi-variable approach is very useful when the history matching parameters are
dependent but each selected region for updating is independent of others.

Key words: reservoir simulation; automatic history matching; parameter updating scheme; time-lapse (4D) history matching; misfit

Introduction which can be considered as gradient based, deterministic or

History matching of oil and gas reservoirs is one example
of difficult inverse problems and finding an appropriate solu-
tion is very important in the oil industry. In this problem usu-
ally we are faced with updating the reservoir simulation
model properties for most of the cells by adjusting the match
to the observed production data. Because of the high number
of unknowns in the model it is very difficult to deal with this
problem manually and researchers have tried to introduce a
range of Automatic History Methods (AHM) to speed up so-
lution of the problem [16].
The AHM methods can be broken into several elements in-
cluding: 1) the optimization algorithm, 2) parameterization of
the simulation model, 3) identification of the important prop-
erties that require updating, which is a combination of uncer-
tainty and importance of the effect on flow 4) definition of the
objective function. The main goals for the whole process are
to reduce the number of unknowns and to make the inversion
process faster and cheaper. In other words efficiency should
be maximized. On top of that, for application of AHM meth-
ods to field data we need to honor the geological constraints
of the model.
Several inversion algorithms have been developed to date
probabilistic. Gradient based approaches include Gauss-
Newton [7] and steepest descent and these require the deriva-
tive of the misfit with respect to the updated properties. Sto-
chastic approaches such as genetic algorithms [8], simulated
annealing [9], Neighborhood algorithm (NA) [10] are global
optimization methods which are based on random search in
parameter spaces without computation of gradients of the
misfit. Probabilistic methods such as the Randomized maxi-
mum likelihood and Ensemble Kalman Filter [11] are more
closely tied to the uncertainty statistics of the updated pa-
rameters and resulting models.
Parameterization methods are used to reduce the number of
unknowns while keeping the geological information of the
model. An example of these methods is gradzone [12] which is
based on calculating the gradient of parameters in a zone
rather than a cell. Alternatively the pilot point method [13] is
based on updating the reservoir properties locally and propa-
gating the properties around changes with Kriging. The grad-
ual deformation method [1416] which is gradually updates the
reservoir model by keeping the geostatistical constraints is
another example in this category.
For appropriate identification of where models should be
changed, researchers have used streamline simulation.

Received date: 27 Jan. 2012; Revised date: 25 Mar. 2012.

* Corresponding author. E-mail: alireza.kazemi@pet.hw.ac.uk
Copyright 2012, Research Institute of Petroleum Exploration and Development, PetroChina. Published by Elsevier BV. All rights reserved.
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
Streamlines were used to help to find where fluids flow so that
these regions can then be modified appropriately [1720].
In this work, we choose to use a stochastic approach. This
approach gives a range of model outcomes in contrast to the
gradient based methods which produce a single solution. For
reducing the number of unknowns we introduced a method
which was based on the guide of streamlines to appropriately
identify the optimal area of the reservoir for updating [20]. For
keeping the geostatistical constraint of the model we used the
pilot point method with Kriging which also helps us again to
reduce the number of unknowns.
We believe that in addition to the above combination of
AHM elements, there is one more that should be considered.
This is the way we combine history matching variables during
the updating (which we called parameter updating scheme). In
this work we mainly focus on identifying an appropriate pa-
rameter updating scheme to for the Nelson field in the North
Sea.
1 General strategy for history matching
In order to have success in solving the history matching in-
verse problem, an appropriate strategy needs to be considered.
This strategy involves analyzing the available data such as
geological information or fluid flow patterns in the reservoir,
studying the sensitivity of the effect of selected parameters on
fluid movement, combining different variables in the history
matching study in order to keep the geological features and to
control the fluid displacement, and finally analysis of the
output of simulations so that we can decide to accept or reject
the new reservoir models. Fig. 1 shows a complete workflow
of history matching of a reservoir.
The first stage in a history matching strategy is likely to be
selection of reservoir variables that require modification.
Also we need to know where and how these variables need to
Fig. 1 Reservoir history matching workflow
 350 
be updated in the reservoir (Step 1 in Fig. 1). Generally for
any history matching study, the reservoir parameters which
include, their location where modifications are made, must
then be chosen using the available knowledge for the uncer-
tainty of various data in the reservoir, the geological under-
standing of the reservoir and also change to the fluid move-
ment that can be expected. In this study the guide for choosing
the reservoir properties initially is mainly based on the geol-
ogy of the reservoir. The challenge in this study (Nelson field)
was to update shale volumes, their distributions and effect on
flow by updating three variables net:gross (NTG), horizontal
(Kh) and vertical permeability (Kv). Multipliers are used to
change those variables and a reasonable range needed to be
chosen based on prior information. For selecting the updating
regions in the reservoir, the idea in this study is that by using
streamlines. Streamlines represent the path of fluid displace-
ment in the reservoir; therefore appropriate selection of re-
gions to update can be made. This idea would be suitable for
the any active region of the reservoir.
For appropriate updating of the reservoir, various parameter
updating schemes are introduced (Step 2 in Fig. 1). However,
certainly there are some limitations and drawbacks for each
scheme that also will be discussed in this paper. In the study,
because of the limitations for running a huge number of
simulation models for high dimensional problems, it was con-
sidered that the maximum number of models should be 4 000
for a given history matching run. Various runs are then com-
pared.
As part of using an AHM workflow (Step 3 in Fig. 1) the
aim was to reduce the mismatch between observed and mod-
eled data by updating the parameters in the reservoir and ap-
plying data analysis (Step 4 in Fig. 1). After analyzing the
parameters, and also the reduction of misfit, there were two
options. If the results were satisfactory and the parameters
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
converged to a specific value, we went to the final steps of the
process (Step 5 in Fig. 1) and by combining the best results an
ensemble of best reservoir models was obtained. Otherwise,
the process returned to the beginning of the loop and chose
another region for reservoir updating, changing the parameters
or modifying the range of the parameters for sampling.
2 Nelson field
The Nelson field (Fig. 2) is an undersaturated oil field lo-
cated in blocks 22/11, 22/6a and 22/12a in the UK Central
North Sea. The production begin in 1994 and 27 production
wells were drilled up to 2000. The original oil in place was
estimated at 1.26108 m3 and up to the end of 2000, 0.47108
m3 has been produced from the field [21].
The reservoir sands in Nelson are turbidities with excellent
reservoir quality with average net:gross of 70%, average po-
rosity of 23% and permeability ranging from 200103
1 700103 m2. Three distinct intervals have been identified
in Nelson separated mainly by shale. Each interval consists of
channel complexes that were gradually filled with sands
through turbidity events. Then because of erosions there are
smaller channels and there is a high chance of shale between
these smaller channels [22]. The result is that we most likely
have some combination of sand and shale layers. The shales
may be discontinuous due to erosion while shale drapes of
smaller channels may lead to reduced horizontal permeability.
Shale content is considered to be a major uncertainty in Nel-
son, therefore, and its volume and distribution affect net:gross
and horizontal and vertical permeabilities. We therefore up-
date these variables but first we must assess what combina-
tions of change might mean. There are 8 combinations pre-
sented where the variables are increased or decreased and, we
interpret these geologically. For example, for the case where
all variables are increased via history matching we could ex-
plain this we modify a very shaly region (indicated by many
black lines in Fig. 3). Increased NTG would represent either
thicker sands, thinner shale deposits or more erosion. In-
creased horizontal permeability would represent less draping
of shale and increased vertical permeability would reflect
Fig. 2 Location map of the Nelson field
 351 
Fig. 3 Geological interpretation of the changes of net:gross or
permeability via history matching
more erosion. Fig. 3 then shows various combinations of pa-
rameter changes schematically. Because each variable can
reflect different conditions separately, treat each independ-
ently rather than use a relationship.
Simulations are carried out in this study, the historical liq-
uid rates as well controls. In order to control the bottom hole
pressure of the wells we used a pressure minimum to make
sure that pressure does not reach to bubble point as no gas
production has been observed other than that dissolved in the
oil. We history matched from 1994 to 2000 and to check the
quality of results 3 years forecasting was carried out using
data up to 2003. For simplicity the 6 production wells with the
highest misfits were targeted. The regions near these wells
were then modified as identified by guide of streamline simu-
lation [20]. In each region 3 parameters are used. We allow an
order of magnitude variation of each parameter either side of
the base case.
The operator of the field supplied a reservoir simulation
model which contains over 500 000 active cells each ap-
proximately 75 m75 m8 m. The reservoir contains several
faults but these faults have a good connectivity and we do not
need to modify their flow properties. The finite difference
model took between 23 h to run on a 3.4 GHz PC while
streamline simulation, with a suitable choice of streamline
density and time step for changing pressure, took about 8
minutes.
3 Parameter updating schemes
The main reason for introducing different parameter updat-
ing schemes, therefore, was because of the lack of computing
power on the one hand and while on the other, the aim was to
retain the geological features of the reservoir by updating the
right parameters in the right locations. In the updating ap-
proaches that were used in this study, various variables were
combined in the model during the AHM process in various
combinations. This was done globally for the all updated re-
gions or locally for each individual region. The global term
here means that parameters were modified for all the selected
regions in the reservoir. Within each updated region, an ap-
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
propriate localized geostatistical method was used to
smoothly propagate the changes to the variables in the reser-
voir. The geostatistical approach used was the pilot point with
Kriging [13].
3.1 Global single variable approach (GSV)
In this approach, just one variable was changed at a time.
This scheme is suitable for those cases where history match-
ing parameters do not have interactions such that they are
independent. As an example, if the history matching problem
is solved by first matching reservoir pressure and then flow
rates (as is often done manually) we can focus on separate
parameters. For example by updating aquifer properties (as a
single variable) or field average permeability through history
matching we can first match the reservoir pressure response.
Then by updating appropriate reservoir parameter such as
permeability variations or relative permeability we can match
liquid production (usually water cut) from the reservoir.
In this study the GSV method was applied to properties
such as net:gross which was changed at all locations selected
for updating in the reservoir while permeabilities are fixed to
capture some of the regional parameter interactions. The same
procedure was then repeated for other variables (fixed vari-
ables were set as the base reservoir model). For this study we
are dealing with three reservoir variables. Updating net:gross
alone can have an important influence on history matching by
increasing the amount of pore volume in the reservoir.
Horizontal permeability updating alone can have an influence
of lateral water displacement in the reservoir which can be
important to match the well performance and vertical perme-
ability has a reasonable effect on water displacement from the
bottom of the reservoir to the producers.
However, there is often some relationship between various
petrophysical properties of the reservoir and changing one
parameter may require updating the other parameters too.
Therefore, there is a small probability that this scheme will
work for the Nelson field.
3.2 Regional multi-variable approach (RMV)
This updating scheme is suitable for a field where there are
dependencies between selected parameters (such as in the
Nelson field) but we can also separate the reservoir into dif-
ferent regions with a low dependency and therefore history
matching can performed in different regions. Here, regions
were identified within which parameter interactions might
occur. Various bases can be used to choose the regions in the
model such as separating the reservoir based on different
geological intervals, separating by reservoir and non-reservoir
sections or using geological bodies such as channels to dis-
tinguish between the regions. In Nelson two separate regions
were identified with the help of streamlines. A search was
performed for all parameters within a region simultaneously
and one region was searched at a time.
 352 
3.3 Local multi-variable approach (LMV)
The name of LMV comes from the fact that history match-
ing parameters will be updated locally (i.e. over a few tens of
grid cells) in the reservoir. This updating approach can be
applied in cases where multiple history matching parameters
affect the misfit locally but where there are many small re-
gions that are independent. In this approach the parameters at
every locality will be modified simultaneously in order to
capture local interactions only. The difference between this
approach and RMV is that, here, the selected regions for up-
dating are much smaller and specifically relate to the drainage
area of a well. Very precise locations were needed to make
sure that there was negligible interaction. All parameters at
each locality were updated and one locality is modified at a
time though with a suitable algorithm all localities could be
updated simultaneously.
4 Automatic history matching workflow
History matching method was based on an automatic work-
flow presented as step 3 in Fig. 1 and is illustrated in Fig. 3.
New models are generated as perturbations about the base
model in regions suggested by streamlines. We use the pilot
point method with Kriging to control updates. Pilot points are
used to directly control where changes are made to properties
such as permeability and net:gross. These changes are inter-
polated laterally using Kriging and applied uniformly in the
vertical direction within a geological interval [23-24]. In order to
reduce the number of unknowns and to spread changes more
smoothly, a set of pilot points may be grouped so that changes
are applied all in the same way. This approach is often called
the master pilot point approach [13].
Modifications at the pilot points are initially chosen ran-
domly and a number of new models are generated. We simu-
late each model by using streamlines which is appropriate for
the field we are studying. In other fields finite difference
simulation may be required. We compare predicted production
data with the observed data using an objective function.
(Qobs  Qcal ) 2
M (1)
VQ
2
We also use the normalized misfit to relate the misfit to the
misfit of the noise (i.e. the misfit that would be calculated if
all parameter errors could be removed and only the remaining
errors in the data are obtained). In the above formulation, the
misfit of the noise is actually just the number of measure-
ments so the normalized misfit is:
J=M/N (2)
where N is the number of measurements.
At the end of the loop we use the Neighborhood Algorithm
(NA) as a stochastic quasi-global optimization algorithm to
select new parameter values automatically. We then iterate
through the loop until the objective function is minimized
sufficiently.
The NA is a stochastic sampling algorithm introduced by
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361

Sambridge in 1999 [11] originally for earthquake seismology.

Using this algorithm we explore the multidimensional pa-
rameter hyperspace using the geometrical constructs known as
Voronoi cells. The process is initialized by generating ni mod-
els using quasi or pseudo-random methods. In each iteration
that follows the best nr models are selected and their
neighbourhoods are identified by construction of Voronoi cells
in the parameter space. ns new models are generated such that
ns:nr models are located within the neighbourhoods of the
each of the best nr models. One drawback of this algorithm is
that for complex misfit surfaces, we require 2nd models ini-
tially to properly separate out neighbourhoods [11].

5 Data analysis
Fig. 4 shows four typical outcomes of the history matching
process and the figure is used to illustrate the results. 27 pa-
rameters were defined for the updating approaches. In Out-
come 1 (Fig. 4a) the parameter converges to a specific value
leading to a lowest misfit and we would hope for this every
time. Outcome 2 (Fig. 4b) shows that the parameter did not
converge to any particular value because in the sampled range
there is no influence on misfit. This may be because the com-
bination of this parameter with other parameters does not
produce a model with lower misfit. Technically, this pa-
rameter does not have an important influence on reduction of Fig. 4 Four typical outcomes of history matching
misfit value. We might find influence if we also change other
parameters in a different range, however. Outcome 3 (Fig. 4c) 6 History matching results
means that more than one value of this parameter leads a
The degree of change to each of the reservoir properties
minimum value. This kind of behavior makes the analysis of
was also analyzed by the degree of improvement in well pro-
history matching more difficult in the choice of models.
duction misfit. The ability of the new reservoir models to
In Outcome 4 (Fig. 4d), the parameter approaches the limits
forecast future liquid production was examined.
that were set at the start of AHM. In this case, in order to
make sure that the conceptual geological model is honored, 6.1 Global single variable approach (GSV)
there is a need to revisit the geological description of the res-
In this approach the goal was to improve one variable at a
ervoir. If increasing the limit of change for the parameters
time using the history matching workflow. The history
makes a big change in the reservoir that is not valid, geologi-
matching was performed to update this reservoir property.
cally, we should not change the range. Otherwise it is possible
to make the range bigger and perform the history matching Then the same procedure was repeated for the other two vari-
again to get a better result. Fig. 4e shows a probability curve ables (vertical permeability and net:gross) and the results were
obtained by Markov Chain Monte Carlo resampling of the analyzed. In this study there were just 9 regions making the
parameter space [25]. The trend for a probability distribution problem 9 dimensional.
curve is compatible with each outcome. For example when The result of history matching for each variable is that the
there is an outcome as in Fig. 4a a probability distribution sum of misfits of the oil and water for the targeted wells de-
curve also shows one peak. This corresponds to the most creased (Fig. 5). Horizontal and vertical permeability had
likely parameter value observed during the history matching more effect on the misfit value compared to net:gross. From
process to give the minimum misfit. this figure it can be concluded that net:gross modification will
After analyzing the parameters and also the reduction of reduce the misfit value by around 50%. Also, changing verti-
mismatch we have two options. If the result is satisfactory, cal and horizontal permeability reduced the misfit by 70% and
and the parameters converge to a specific value we go to the 75% respectively. Of the 27 parameters 13 approached the
final step by combining the parameters. An ensemble of best upper limit and 5 the lower limit. It was decided that for those
reservoir models is obtained. Otherwise we need to back to reaching the upper limit, the limit should be increased to 1.3
the beginning of the loop and find alternative ways of updat- (which is equal to a factor 20 multiplier on a linear scale, al-
ing, either changing the parameters or changing the intervals though changes are highly localized) and the lower limit was
of parameter multipliers. changed to2 for those parameters that approached the
 353 
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
Fig. 5 Sum of misfits of oil and water rates for the targeted
wells versus model index from AHM by changing one variable at
a time
Fig. 6 Sum of misfit of the oil and water rates for the targeted
wells versus model index for AHM result of changing one pa-
rameter type at a time
previous lower limit. Increasing the upper limit to more than
1.3 was considered to be geologically invalid. History match-
ing was then repeated with the new limits and all other pa-
rameters the same.
Fig. 6 shows the reduction of the sum of production misfits
for the 6 wells after modification of each parameter range.
From this figure there was now 50%, 80% and 75% reduction
in sum of misfits of the oil and water for net:gross, Kh and Kv
respectively. Comparing with the first step of history match-
ing there is a further 10% reduction in the misfit by extending
the range of Kh limits.
Probability distributions of each parameter (Fig. 7) show
for three parameter representing two different regions of the
Fig. 7 Probability of different parameters for two different regions
 354 
reservoir were obtained by resampling the misfits using
Monte Carlo Markov Chain integration [26]. Distinct trends
were observed. For example the Kh line in Fig. 7a shows that
the highest probability of the parameter at the parameter limit.
This kind of trend is consistent with Fig. 4d. Other parameters
have peaks within the limits.
The net:gross and Kv lines in Fig. 7a show a wide bell
shaped probability distribution. This trend is consistent with
Fig. 4b. The trend for Kv in Fig. 7b shows a narrow probabil-
ity curve for this parameter which means that a specific value
of this parameter was selected repeatedly during generation of
reservoir models and this behaviour is similar to the Fig. 4a.
The Kh trend in Fig. 7b has two peaks in the probability dis-
tribution indicating two possible solutions for this parameter.
One of the solutions could be a local minimum in the misfit.
Such a result indicates the problem of non-uniqueness in his-
tory matching more clearly. In this study it can be seen that
even after widening the limits there was a high probability of
obtaining a parameter value near the new limits of the interval
(Kh in Fig. 7a). As discussed previously this parameter interval
could not be made bigger because it changes the conceptual
geological model too much. It may be concluded also that the
selected parameters that were changed for that specific region
were not sufficient and others should be found.
At the end of each history matching study we require a
model (or a set of models) including the most appropriate
changes that were made to the reservoir parameters. On the
other hand as previously discussed, for the field in this study,
the updated reservoir parameters may be inter-dependent and
the history matching results need to be analysed after com-
bining the resulting parameters. For this purpose all models
were sorted for the history matching runs by each variable
based on the decrease of the sum of misfits of the oil and wa-
ter rates for the targeted wells. Then, the parameters of the 10
models with lowest misfits were chosen and combined to get
one reservoir model with updated permeabilities and net:gross
for each rank (Fig. 8).
The functions NTGi(u), Khi(u) and Kvi(u) represent the up-
dated reservoir variables net:gross, horizontal permeability
and vertical permeability respectively (as a function of loca-
tion vector, u) after history matching and i=1,2,3,..10 to de-
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361

Fig. 8 Procedure for generating 10 final simulation models

tically and then the aquifer support was changed to be bottom

Fig. 9 Sum of misfits of the oil and water rates for the targeted
wells in the history and forecasting periods for the 10 generated
models
drive. Also, decreased horizontal permeability strengthened
bottom drive compared to edge drive.
However, there are still some general advantages for the
GSV approach. For example for some history matching cases
this approach can be use for investigating the effect of differ-
ent types of variables that may be updated. Then by ignoring
the less important variables and combining the remainder, it is
possible to efficiently modify the reservoir. On the other hand
another application of this approach would be for the cases
where parameters are independent. For example if in a case,
the fault transmissibilities and aquifer properties are updated
then with the GSV approach, we can modify the aquifer prop-
erties and then fault transmissibility.

note the order of misfit from best to worst (lowest is for i=1). 6.2 Regional multi-variable approach (RMV)
Thus, the final simulation models were generated by combin- In this approach, the reservoir was divided into two parts
ing the updated reservoir parameter such that: with the idea that there was a very low interaction between
GSVFinal 1=f (NTG1(u), Kh1(u), Kv1(u)) (3) reservoir parameters between these two regions. There are
To compare between these final models, the sum of misfits three wells in the Edge of the reservoir (Fig. 10), where four
of the oil and water rates for the targeted wells can be seen in locations were identified for updating; there were three in the
Fig. 9 for the history matching period, and for forecasting. All first interval and one in the second interval. Also, there are
models were sorted by misfit and then they were compared three wells in the Centre of the reservoir consisting of 5 loca-
with the base case model. It can be observed that the combi- tions of pilot points; three in the first interval and two in the
nation of the best values for each parameter after history second interval. Based on this division we created 12 parame-
matching did not improve the well misfit that much and in
forecasting was even worse as the misfit value was higher
than the base model. This result shows the importance of con-
sidering different combinations of parameters while history
matching.
The Nelson field is complex changes to properties alter the
fluid flow. For example a region near the aquifer with a very
low vertical permeability and net:gross, represented a quite
shaly region where the shales were very well connected hori-
zontally and acted as a vertical flow barrier. If we only in-
creased net:gross we effectively added more sand but if verti-
cal permeability stayed low the remaining shale was still well
connected. This increased the edge drive from the aquifer. If
we increased net:gross and vertical permeability together we
created a sandier region which was very well connected ver- Fig. 10 Identification of the regions used in this study
 355 
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
ters in the Edge section and 15 parameters in the Centre sec-
tion.
For Edge study 2 700 models were used and for Centre
study 4 000. Obviously, in the Centre, more models were used
because it is a 15 dimensional problem. Fig. 11 shows the
total misfit for the targeted wells during history matching
workflow. In this work the Edge and Centre regions contrib-
uted three wells each to the total misfit. According to Fig. 11,
the misfit decreased by 85% in total. In this case, the parame-
ters evolved as in Fig. 4b mainly because of the dimension of
problem. We did not generate enough reservoir models and
convergence was not observed.
Fig. 12 shows the base model and the ten best models for
the history matching period (up to 2000) and the misfits com-
Fig. 11 Sum of misfits of the oil and water rates for the targeted
wells in each region versus models during history matching in
Edge and Centre of reservoir
Fig. 12 Sum of misfits of the oil and water rates for the best 10
models in the RMV case
 356 
paring differences between 2000 and 2003 indicating the de-
gree of improvement in the forecasting period. A very good
reduction in the misfits was observed for the best models and
also, in the forecasting period, the misfits decreased. Model
10 is the best model for both history matching and forecasting
periods.
6.3 Local multi-variable approach (LMV)
In this approach, the aim was to update the reservoir model
in 6 different regions near the wells, one by one, by changing
all three variables. Three of the updated regions were in the
first interval such that the sub-problem was 3 dimensional in
these three cases. The others were located in both intervals
and the sub-problem was 6 dimensional. In this case, only the
misfit of the specific well located in the region of study was
calculated during updating. The reduction of the well produc-
tion misfit through history matching for each well was ranged
from 82% to 95% (Fig. 13). All parameters converged as in
Figure 4a. The parameter multiplier range did not need to be
modified in this case.
Similar to GSV and RMV for this scheme also we gener-
ated 10 models by using updated reservoir parameters of each
local history matching study. Imagine the abbreviation Local-
ityji (j=1, 2, , 6) for history matching result of each locality
and i=1, 2, , 10 to show best history matched model in
terms of misfit value sorted from high misfit to low misfit
(lowest is for i=10) in following figures. Then the final mod-
els were generated by combining the updated reservoir pa-
rameter such that:
LMVFinal1 ( NTG, K h , K v ) Locality 1 1 
( NTG, K h , K v ) Locality 2 1  "  ( NTG, K h , K v ) Locality 6 1
#
LMVFinal10 ( NTG, K h , K v ) Locality 1 10 
( NTG, K h , K v ) Locality 2 10  "  ( NTG, K h , K v ) Locality 6 10
As in the previous studies the misfit of the 10 final models
(LMVFinal) was again selected. Fig. 14 shows the sum of mis-
fits of the oil and water for the targeted wells for each model
in the matching and the forecasting periods. Between these
models, Model 10 was the best reservoir model because of the
low misfit value.
7 Comparison of the updating approaches
There are 4 basic metrics used for comparison: (1) the CPU
time used to get a reasonable result; (2) the ease of analysis
and the understanding of the result from each scheme; (3) the
scheme with most reduced misfit and also least damage to the
prior geological model of the reservoir; (4) the improvement
to the forecast.
Table 1 shows the number of history matching steps and
models needed in each case. In terms of CPU time, we need at
least 137 processor-days using the global single variable
scheme compared to 10 days for the local multi-variable
scheme. Further, generating more models with any given
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361

Fig. 13 Well water production misfit versus model index for 6 wells in different part of reservoir

Table 1 Number of models used in history matching and total

CPU time
Updating History mat- Total number CPU
Number of models
scheme ching steps of models time/d
4 104 (NTG); 4 104
GSV 2 24 626 137
(Kh); 4 104 (Kv)
2 720 (Edge);
RMV 1 6 790 37.5
4 070 (Centre)
198 (3 Dimension);
LMV 1 1770 10
1 572 (6 Dimension)

Table 2 Parameter outcome for each scheme

Fig. 14 Sum of misfits of the oil and water rates for the best 10 Number of parameters
Outcome
models in the LMV case GSV RMV LMV
Fig. 4a 9 0 19
scheme means that we need to spend more time in analysis
Fig. 4b 6 27 5
and this means more human interaction. We do not have a
Fig. 4c 5 0 3
metric for that but we need to keep it in mind. Table 2 sum-
Fig. 4d 7 0 0
marizes the number of parameters with various behaviors (as
in indicated Fig. 4). This table shows that for the RMV comparison is the misfit value for the best 10 models after
scheme there is no clear behavior about how the parameters history matching as shown in Fig. 15. It can be seen that the
evolve and all 27 parameters behave as in Fig. 4b. In the GSV GSV parameter updating scheme less improvement than the
scheme, convergence was obtained for 9 parameters while 7 other two approaches, which were very much acceptable.
tended to the limits (even an increase). From a geological Fig. 16 crossplots the reduction of misfit for history match-
perspective, it was not possible to allow properties to increase ing versus the forecast period. For the GSV case, there was
by more than a factor of 20. However in the LMV case, all 19 some reduction of the misfit in the matching period but the
parameters converged making analysis easier. forecast was actually poorer. For the RMV and LMV cases,
After analysis of the parameters, the history matching re- the best 10 models gave a good forecast of the targeted wells.
sults between the three schemes were compared. The first There was also a good correlation in the results.
 357 
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
Fig. 15 Sum of misfits of the oil and water rates for the best 10
models of each case
Fig. 16 Forecasting misfit reduction versus matching for the
best 10 models of each case
For the best model (defined as the best combination of his-
tory and forecast misfit reduction) for each scheme, we
cross-plotted individual oil production misfits of each well
Fig. 18 Best parameter multipliers applied to update the base model in various cases
 358 
compare with the base model (Fig. 17). The GSV case gave
misfits that were close to the base case and there were high
misfit values for both periods. In each of the RMV and LMV
cases there was a big shift of the models toward zero which
means that the well oil production misfit was reduced for both
periods.
One important outcome of history matching was the
changes to the representation of fluid movement in the model.
Fig. 18 shows the multipliers of each variable for the best
reservoir model. The horizontal permeability was increased
while net:gross and vertical permeability were increased in
some locations and or decreased elsewhere. For example in
Region 1 (indicated by the arrow in Fig. 18) the lateral sweep
of oil towards the well was increased following the increased
horizontal permeability along with decreased net:gross and
vertical permeability. This result was obtained from the RMV
and LMV schemes. In GSV, however, increased vertical per-
meability was obtained.
Region 2 in Fig. 18 is close to an injector and the lateral
sweep was again increased as in Region 1 but here there was
also reduction of shale volume by increasing net:gross. For
Fig. 17 Oil production misfit in forecasting period versus his-
tory period for the best model from each scheme
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
this region almost the same changes were required for all up-
dating schemes. Region 3, close to the edge of the reservoir as
well as an injector, again saw increased lateral sweep of oil by
updating the properties. Therefore, it can be said that that the
injector has an important impact on the production well.
Again, for this region there was a similar update for three
cases. In the fourth region, which was in the centre of the res-
ervoir, the only apparent route for oil displacement was from
the bottom aquifer. For the RMV and LMV approaches, the
fluid displacement was controlled by increasing the ratio of
vertical to horizontal permeability. In RMV both permeabili-
ties were increased but more for the vertical direction and in
LMV the vertical permeability was kept unchanged while the
horizontal permeability decreased.
One general conclusion based on Fig. 18 is that even for
small reductions of the misfit value in the GSV case, the
process of updating the reservoir was very similar to other
approaches except that usually the degree of change was very
large for this case. Therefore, it can be said that by consider-
ing only one variable at a time, the fluid movement was al-
tered to increase or decrease lateral movement compared to
other cases but updating was as not as efficient. The efficiency
of fluid displacement could be increased by considering the
change to all the parameters at the same time (RMV and LMV
methods).
By adding more wells into the history matching study, the
number of updated regions would be increased as well. More
reservoir parameters are changed thus increasing the dimen-
sion of the problem. As discussed several times during this
paper, increasing the dimension of the problem requires
greater searching in the parameter space and this means sig-
nificantly increasing the number of simulations. From this
point of view the LMV approach would be more suitable than
RMV and GSV, because it needs fewer simulation models.
In summary, the GSV is suitable for cases where the effect
of properties chosen for updating are independent, such as
fault transmissibilities, aquifer properties etc. The RMV is
suitable for the cases where there is strong dependency be-
tween properties chosen for updating. Also there were wells
very close together with strong interaction. In that case, up-
dating the reservoir parameters by regions influenced all wells.
The LMV method is suitable for cases where there are local
dependencies between the properties selected for updating but
the regions to update are independent.
8 Application of LMV method for production and
seismic history matching of Nelson field
In addition to the production data, 4D seismic also inte-
grated into the objective function. Therefore the misfit func-
tion is updated as below
( Sobs  Scal ) (Qobs  Qcal )
2 2
M V s2

V Q2
(4)
The observed 4D seismic in Nelson is phase shifted ampli-
tude (is a type of colored inversion) maps. However the phase
 359 
shifted amplitude can be considered as pseudo impedance
because there is zero crossing trace for interface in both types
of data. In each reservoir interval the value of phase shifted
amplitude could be equivalent to the impedance value. Al-
though the outcome of colored inversion is equivalent to im-
pedance, based on the inversion process the unit of product is
the same as seismic amplitude. In order to quantitative inte-
gration of colored inversion into history matching loop we
need some kind of calibration and normalization in order to
have a unit equivalent to the unit of synthetic acoustic imped-
ance data. In Nelson for phase shifted amplitude the root
mean square of each sample interval was calculated and
therefore a 2D map was generated for each reservoir interval
and for the time differences. Fig. 19 shows as an example map
for the top reservoir interval.
Following the strategy used in this study, reservoir proper-
ties were updated in the model. Predictions of seismic imped-
ances [24,25] are made by first calculating the saturated bulk and
shear moduli for each simulation cell using output from the
simulator. From this the p-wave modulus is calculated and
upscaled vertically using Backus [27] to give a single value for
the reservoir interval. Bulk density is similarly calculated and
then the bulk impedance is obtained from:
I mod IJ  U IJK ! K  ( ksat IJK  4 P IJK / 3) 1 ! K (5)
where, < >K is the depth average weighted by volume.
Seismic attributes were obtained as an equivalent pseudo-im-
pedance property for each bin in the seismic cube. These data
were then averaged over each simulator cell to give a value
that can be compared to the predicted data using
I obs IJ  I obs ij ! xy (6)
where, < >xy is a really average upscaling from the seismic
to the simulator grid.
We also calculate the standard deviation of the observed
seismic within each simulation cell as a measure of the vari-
ability. A similar misfit is used for production data summing
Fig. 19 The maps of differences of phase shifted amplitude
(colored inversion) for top reservoir interval in 19902000
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
over time series data of oil and water production rates. The
production and seismic misfits are added to give a total misfit.
In this study we have targeted the 13 worst matching wells
with one master pilot point per well. These wells make up 84
per cent of the total production misfit. 7 of these wells are
completed in the top geological interval so we do not change
properties at those locations in the intervals beneath. In this
case each history matching problem is three dimensional. The
other 6 remaining wells that we focus on are completed in
both oil filled intervals where we update properties. The
problem is therefore six dimensional for these regions. Over-
all, the problem is 57 dimensional by using the local
multi-variable approach is a combination of 7 three dimen-
sional problems and 6 six dimensional problems. After com-
bination of updated reservoir parameters we again selected the
best 10 history matched models.
The history matching result is investigated for its capability
to predict the future behavior of the reservoir (i.e. beyond the
history matching period to forecast from 2000 up to 2003).
For this purpose the well production data between 2000 and
2003 were used to measure the forecast accuracy of the best
10 models from each case. We know that because the history
matching is non-unique we do not expect the forecasts to be
the same. From Fig. 20 it is clear that we successfully update
the reservoir properties by reduction of misfit with 50% and
30% in history matching and forecasting periods. Fig. 21
shows the time-lapse (4D) signature prediction of the best
history matched model. Comparing each model with the base
model (Fig. 21b) a general improvement of seismic in the
middle of the reservoir was observed.
9 Conclusions
Three updating schemes were introduced for better history
Fig. 21 Seismic impedances data of top reservoir interval
Nomenclature
Mmisfit;
<>misfit function;
'Qthe difference between observed and calculated production
 360 
matching. Global single-variable approach is only a suitable
parameter updating scheme for cases where the history
matching parameters are independent. For reservoirs with a
high density and strong dependency of the wells, it is better
that the regional multi-variable approach is used for updating
the parameters. Local multi-variable approach is very useful
when the history matching parameters are dependent but each
selected region for updating is independent of others. They
were all applied in the Nelson field, and the local multi-varia-
ble scheme was found to be the best. By using local scheme
the number of simulation models needed for history matching
reduced by 90% and 70% compared to global and regional
parameter updating schemes. For selected wells in history
matching, we improved the misfits of oil and water rates were
reduced by 90% and 60% for history and prediction periods
respectively for local updating scheme. In seismic and pro-
duction history matching, for all wells in the reservoir, we
improve the misfits of oil and water rates by 50% and 30% in
history and prediction period respectively. A general im-
provement of seismic in the middle of the reservoir was ob-
served.
Fig. 20 Reduction of total production misfit in forecasting ver-
sus matching periods
data;
|C |the covariance of observed data;
Sobsobserved seismic data;
Scalcalculated seismic data;
sstandard deviation of observed seismic data, m3/d;
 KAZEMI Alireza et al. / Petroleum Exploration and Development, 2012, 39(3): 349361
Qobsobserved oil and water rate, m3/d;
Qcalcalculated oil and water rate, m3/d;
Qstandard deviation of observed seismic data, m3/d;
Jnormalized misfit;
Nnumber of measurements of observed data;
nnumber of variables;
Imodsimulated acoustic impedance, (Pas)/m3;
shear moduli, Pa;
3
density, kg/m ;
ksatbulk modulus, Pa;
I, J, Kx,y,z indices on simulation grid;
Iobsobserved acoustic impedance, (Pas)/m3;
x, ydirection on seismic grid.
References
[1] Watson A T, Lee W J. A new algorithm for automatic history
matching production data. SPE 113479, 1986.
[2] Ouenes A, Brefort B, Meunier G, et al. A new algorithm for
automatic history matching: Application of simulated
annealing method (SAM) to reservoir inverse modeling. SPE
26297, 1993.
[3] Tan T B. A computationally efficient Gauss-Newton method
for automatic history matching. SPE 29100, 1995.
[4] Bissell R C, Dubrule O, Lamy P, et al. Combining
geostatistical modelling with gradient information for history
matching: The pilot point method. SPE 38730, 1997.
[5] Portella R C M, Prais F. Use of automatic history matching
and geostatistical simulation to improve production forecast.
SPE 53976, 1999.
[6] Zhang F, Skjervheim J A, Reynolds A C, et al. Automatic
history matching in a Bayesian framework, example
applications. SPE Reservoir Evaluation & Engineering, 2005,
8(3): 214223.
[7] Fletcher R. Practical methods of optimisation. New Jersey:
Wiley, 1987.
[8] Holland J H. Adaptation in natural and artificial systems. Ann
Arbor: University of Michigan Press, 1975.
[9] Metropolis N, Rosenbluth A W, Rosenbluth M N, et al. Equa-
tion of state calculations by fast computing machines. Journal
of Chemical Physics, 1953, 21(6): 10871092.
[10] Sambridge M S. Geophysical inversion with a neighbourhood
algorithm(I): Searching a parameter space. Geophysical
Journal International, 1999, 138: 479494.
[11] Evensen G. Sequential data assimilation with a non-linear
quasi-geostrophic model using Monte Carlo methods to fore-
cast error statistics. Journal of Geophysical Research, 1994,
99(C5): 143162.
[12] Aanonsen S I, Aavatsmark I, Barkve T, et al. Effect of scale
 361 
dependent data correlation in an integrated history matching
loop combining production data and 4D seismic data. SPE
79665, 2003.
[13] De Marseily G H, Levendan G, Boucher M, et al. Interpreta-
tion of interface test in a well field using geostatistical tech-
niques to fit the permeability distributions in a reservoir
model// Verly G. Geostatistics for natural resources charac-
terization. Dordrecht: Reidel Publishing Corporation, 1984.
[14] Kretz V, Le Ravalec-Dupin M, Roggero F. An integrated res-
ervoir characterization study matching production data and 4D
seismic. SPE Reservoir Evaluation & Engineering, 2004, 7(2):
116122.
[15] Mezghani M, Fornel A, Langlais V, et al. History matching
and quantitative use of 4D seismic data for an improved res-
ervoir characterization. SPE 90420, 2004.
[16] Roggero F, Ding D Y, Berthet P, et al. Matching of production
history and 4D seismic data: Application to the Girassol Field,
Offshore Angola. SPE 109929, 2007.
[17] Baker R O. Streamline technology: Reservoir history match-
ing and forecasting: Its success, limitations, and future. Jour-
nal of Canadian Petroleum Technology, 2001, 40(4): 2327.
[18] Agarwal B, Blunt M J. A streamline-based method for assisted
history matching applied to an Arabian Gulf field. SPE
Journal, 2004, 9(4): 437449.
[19] Maschio C, Schiozer D J. Assisted history matching using
streamline simulation. Petroleum Science and Technology,
2005, 23(7): 761774.
[20] Kazemi A, Stephen K D. Optimal parameter updating in
assisted history matching of the Nelson field using streamlines
as a guide. SPE 131540, 2010.
[21] UK DTI. UK Monthly Oil Production. http:// www.og.decc.
gov.uk/pprs/full_production.htm.
[22] Stephen K D, Shams A, MacBeth C. Faster seismic history
matching in UKCS reservoir. SPE 107147, 2007.
[23] Stephen K D, Soldo J, Macbeth C, et al. Multiple model seis-
mic and production history matching: A case study. SPE
Journal, 2006, 11(4): 418430.
[24] Stephen K D, Shams A, Macbeth C. Faster seismic history
matching in a United Kingdom continental shelf reservoir.
SPE Reservoir Evaluation & Engineering, 2009, 12(4):
586594.
[25] Sambridge M. Geophysical inversion with a neighbourhood
algorithm: Appraising the ensemble. Geophysical Journal
International, 1999, 138(2): 727746.
[26] Mosegaard K, Sambridge M. Monte Carlo analysis of inverse
problems. Inverse Problems, 2002, 18(3): R29R54.
[27] Backus G E. Long-wave elastic anisotropy produced by hori-
zontal layering. Journal of Geophysical Research, 1962,
67(11): 44274440.